Analytica Chimica Acta 824 (2014) 64–70
Contents lists available at ScienceDirect
Analytica Chimica Acta journal homepage: www.elsevier.com/locate/aca
Mimicking Daphnia magna bioassay performance by an electronic tongue for urban water quality control Dmitry Kirsanov a,b, *, Evgeny Legin b,c , Anatoly Zagrebin d , Natalia Ignatieva d, Vladimir Rybakin d , Andrey Legin a,b a
Laboratory of Chemical Sensors, St. Petersburg State University, St. Petersburg, Russia Laboratory of Artificial Sensor Systems, ITMO University, St. Petersburg, Russia Sensor Systems LLC, St. Petersburg, Russia d Institute of Limnology, Russian Academy of Sciences, St. Petersburg, Russia b c
H I G H L I G H T S
G R A P H I C A L A B S T R A C T
Daphnia magna bioassay can be simulated with multisensor system. Urban water toxicity can be predicted from potentiometric ET data. Independent test set validation confirms statistical significance of the results.
A R T I C L E I N F O
A B S T R A C T
Article history: Received 13 November 2013 Received in revised form 14 March 2014 Accepted 15 March 2014 Available online 20 March 2014
Toxicity is one of the key parameters of water quality in environmental monitoring. However, being evaluated as a response of living beings (as their mobility, fertility, death rate, etc.) to water quality, toxicity can only be assessed with the help of these living beings. This imposes certain restrictions on toxicity bioassay as an analytical method: biotest organisms must be properly bred, fed and kept under strictly regulated conditions and duration of tests can be quite long (up to several days), thus making the whole procedure the prerogative of the limited number of highly specialized laboratories. This report describes an original application of potentiometric multisensor system (electronic tongue) when the set of electrochemical sensors was calibrated against Daphnia magna death rate in order to perform toxicity assessment of urban waters without immediate involvement of living creatures. PRM (partial robust M) and PLS (projections on latent structures) regression models based on the data from this multisensor system allowed for prediction of toxicity of unknown water samples in terms of biotests but in the fast and simple instrumental way. Typical errors of water toxicity predictions were below 20% in terms of Daphnia death rate which can be considered as a good result taking into account the complexity of the task. ã 2014 Elsevier B.V. All rights reserved.
Keywords: Bioassay Water toxicity Multisensor systems Electronic tongue
1. Introduction Water quality and safety is a question of highest importance nowadays. Extensive anthropogenic influence led to massive
* Corresponding author. Tel.: +7 812 328 28 35. E-mail address: [email protected] (D. Kirsanov). http://dx.doi.org/10.1016/j.aca.2014.03.021 0003-2670/ ã 2014 Elsevier B.V. All rights reserved.
hydrosphere pollution all over the world. The concerns of the modern society regarding this issue are reflected in numerous legislative initiatives in this field, such as European Union Water Framework Directive [1], Clean Water Act [2]. The number of analytical techniques and methods suggested for water quality control is really huge and growing. It would be hard to name an analytical method which was not yet suggested for water analysis. Chromatographic, electrochemical, optical, etc. methods can be
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
effectively employed to solve certain particular tasks in this field. One of the integral parameters which are of ultimate importance in ecological water control is its toxicity. Toxicity is understood as a measure of a real harm to the living creatures caused by water sample. There is a special group of methods called bioassay which is intended for direct estimation of water toxicity for biological objects. The general idea behind this type of analysis is to monitor a biological reaction of a biotest object placed in a water sample under study. This reaction is measured in some appropriate way and compared with that of the same biotest object placed in a control sample of pure nontoxic water. Growth rate, mobility, survivorship rate of the test organisms, etc. can be monitored as biological reactions [3]. Various aquatic biota species can be employed for testing, such as e.g. algae, luminescent bacteria, infusoria, daphnia, fish, etc. An obvious advantage of this approach is direct information about biological harm that can be caused to the test object by contaminated water, while most of the other methods based on physicochemical measurements can only provide information on the content of selected chemical substances. This is, of course, useful but not always directly related to the danger posed for living beings, especially when several contaminants are present in the water simultaneously that is quite common. It is important mentioning that sometimes different substances with dissimilar chemical structures can cause the same degree of biological damage in various organisms [4]. Different bioassays and test procedures were suggested and some of them are now accepted as standard analysis legitimized in ISO standards, e.g. Vibrio fischeri (luminescent bacteria) [5], Daphnia magna [6], green algae [7]. It is noteworthy that none of the suggested methods can serve as a universal toxicity assessment instrument because of the various sensitivity patterns of different bioassays. Being an indispensable technique for toxicity assessment bioassay methodology still has a number of drawbacks: its implementation requires highly specialized laboratories; most of the methods are quite slow (usually it takes 24–72 h to get the results); biotest objects must be properly fed and kept under thoroughly regulated conditions. There are numerous publications in literature where attempts were made to overcome these bioassay drawbacks by developing new sensor technologies, see e.g. recent reports [8,9]. The U.S. Environmental Protection Agency (EPA) has certified several Rapid Toxicity Testing Systems (RTTS), like e.g. MicroTox1 [10]. A thorough inspection of the protocols of such methods reveals that all of them require either some biological objects (e.g. freeze-dried bacteria, usually luminescent Vibrio fischeri) or bio/ chemical reagents or quite sophisticated preparative procedures. Besides, some of these methods require highly qualified personnel for operation and in many cases reagent sets include bacteria or living cells, which makes them costly and imply special conditions for their storage. This makes such RTTS and the like methods hardly suitable for continuous autonomous monitoring of water toxicity. Electronic tongue (ET) methodology proved to be useful in complex assessment of various liquid media [11–13]. The operation principle of such systems is quite simple – it suggests the use of an array of cross-sensitive chemical sensors and subsequent processing of unresolved analytical signal from a sensor array by modern multivariate statistic methods [14]. A cross-sensitivity of sensors means that each sensor of the array has its own sensitivity and selectivity pattern and can respond with different response value to various substances of interest in a sample. The resulted signal from all sensors of such array contains information on overall sample composition (according to sensitivity patterns of the sensors) and can be related to certain quality parameters of the samples by means of multivariate data processing both in qualitative and quantitative way. The ET research field is rapidly
65
growing due to a distinct trend in modern analytical chemistry towards fast and simple instruments [15,16]. Our preliminary research [17] with various aquatic organisms (Daphnia magna, Chlorella vulgaris, Paramecium caudatum) has shown that potentiometric multisensor system being preliminary calibrated against bioassay results for a set of samples with known toxicity allows for prediction of toxicity in new, totally unknown water samples in terms of bioassay (e.g. survivorship rate). This procedure does not require complex analytical instrumentation, sophisticated sample preparation and measuring procedures and elaborated biological object maintenance. The idea of the present study was to extend our preliminary studies by a full-scale experiment with sufficient number of samples to explore the opportunities and limitations of ET technology for water toxicity detection. For this purpose we participated in ongoing research initiative devoted to the continuous monitoring of urban water reservoirs in St. Petersburg (Russia). In the framework of this activity regular hydrochemical and toxicity analysis of pond water had been performed, thus we were able to analyze sufficient amount of samples by potentiometric multisensor system. A full access to the results of hydrochemical analysis and toxicity assessment permitted proper statistical validation of the ET predictive performance.
2. Materials and methods 2.1. Samples A series of pond water samples was collected during several experimental sessions in July and August 2012. 29 ponds located in St. Petersburg city area and its closest suburbs were sampled. Water samples from each pond were taken from two horizons – surface layer (0.5 m thick) and near-bottom layer, yielding 58 samples in total. The two samples from different horizons of the same pond were not mixed together and were treated separately in all further studies. Each sample was split into three portions: the first one for classical hydrochemical analysis, the second one for bioassay with Daphnia magna and the third one for electronic tongue measurements. All samples had the same history of storage and transportation before various analyses. Due to diverse location of sampling various pond water samples had various visual appearance, some of them were clear and transparent while the others were turbid, with brownish color and distinct smell. All samples were encoded by a number and a letter: thus 15s stays for the sample taken from the surface (s) of the 15th pond and e.g. 3b stays for the sample taken from the bottom (b) of the 3rd pond. The depth values for the studied ponds varied in the range 1–4.8 m with the average value 2.1 m; traffic conditions in the immediate proximity of the ponds were from heavy to zero (for several ponds located in the park areas). Temperature of the sampled water was in the range 11–24 C with the average value 18.7 C.
2.2. Hydrochemical analysis 15 different hydrochemical parameters were determined for all samples, these were pH, dissolved oxygen, total phosphorus (TP), inorganic phosphorus (IP), electric conductivity, carbon dioxide, ammonium, nitrite, nitrate, total nitrogen, color, chemical oxygen demand (CODCr), biochemical oxygen demand, suspended solids, oil hydrocarbons. The methods of their assessment and their range in the studied samples are presented in Table 1. The determination of these parameters is regulated by normative documents developed by Russian Federal State Budgetary Institution “State Hydrochemical Institute”.
66
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
2.3. Daphnia magna bioassay One- or two-day-old Daphnia magna Straus were employed in 96 h toxicity tests. Ten animals were used for a study in each sample and all studies were performed in three replicate runs. 25 mL of sample was used for each replica. The sample was considered being toxic when 50% or more of the animals died. The fact of death was determined visually using optical microscope. The data from three replicas were averaged for further processing. Resulted toxicity values varied from 0 to 100% with the average 44.4% and the median 35.0%. Due to technical issues bioassay results were obtained only for 50 samples. 2.4. Multisensor system analysis Multisensor system (Electronic Tongue) employed in this study was based on 19 potentiometric cross-sensitive sensors. Our previous trial [17] employed 23 sensors, however during the data processing we found that only 19 sensors have reasonably high regression coefficients in PLS models for prediction of water toxicity in terms of Daphnia magna bioassay while other were useful for infusoria and chlorella prediction. Thus, only these 19 sensors were employed for described experiment. Six of them were poly(vinylchloride) (PVC)-plasticized anion-sensitive sensors based on various anion-exchangers, seven were PVC-plasticized cation-sensitive sensitive sensors with neutral ligands described in Ref. [18], and six sensors were chalcogenide glass electrodes with various RedOx sensitivity patterns. The details on sensor compositions and preparation procedures are widely available in literature, see e.g. [18–20]. Electrochemical measurements were carried out in the following galvanic cell: Cu|Ag|AgCl, KClsat|sample solution|sensor membrane|solid contact|Cu EMF values (sensor potentials) were measured with 0.1 mV precision against the standard silver/silver chloride reference electrode using made-in-house 32-channel digital high impedance voltmeter connected to a PC for data acquisition. All samples were measured in four physically different replicas and the results were averaged for further processing. Replicated measurements of different samples were taken in random order. Water samples were measured as is, without any sample preparation procedures,
dilution, etc. The measurement time in each sample was 3 min and three last readings were averaged and further employed for data processing. After the measurements sensors were washed by several portions of distilled water for 2–3 min. This procedure allowed for 5 mV reproducibility in replicated measurements of the same samples. The measurements in pond water samples with multisensor system were performed in parallel with bioassay measurements in physically different portions of the same pond waters. 2.5. Data processing The results of potentiometric measurements in the samples were combined into matrix form with samples in rows and sensors in columns. Thus, each matrix element was the reading of the sensor k in the sample i. The dimensions of the ET matrix were 58 samples 19 sensors. The same was done for the data of hydrochemical measurements, which yielded the matrix 58 samples 15 parameters. The data from bioassay resulted in a vector 50 samples 1 parameter. We applied several different techniques to relate these two data sets with each other and also to relate multisensor system response with toxicity values produced by Daphnia magna bioassay. Principal component analysis (PCA), canonical correlation analysis (CCA), partial least squares regression (PLS) and partial robust M-regression (PRM) were employed for data analysis. PCA is a widely applied method for data dimensionality reduction and visualization of hidden data structure. The method is based on projections of initial samples from initial multivariate space onto a new coordinate space where new coordinate axes (principal components) are located in the direction of maximum variance and are mutually orthogonal; see e.g. [21] for details. CCA is aimed for the assessment of the correlation between two data sets sharing the same row mode i.e. obtained for the same sample set [22]. To apply CCA we preliminary compressed the ET and hydrochemical data with PCA without significant information loss and after that CCA was run on the resulting PC scores for both data sets. This approach was shown to be effective for ET data processing in wine analysis [23]. PLS is a classical tool for multivariate regression, it finds the regression coefficients vector with a condition of covariance
Table 1 Summary on hydrochemical methods. Parameter
Analytical method and instrumentation
Range min–max (average)
pH, pH units Dissolved oxygen, mg L1 Inorganic phosphorus, mg P L1
Electrometric determination by glass electrode, pH-meter (25 C) Titrimetric determination by Winkler (iodometric method) Colorimetric determination with ammonium molybdate and ascorbic acid, l = 880 nm; spectrophotometer Oxidation with K2S2O8 with subsequent IP determination Conductometric determination in situ (25 C); conductometer Calculative method based on pH and HCO3 data Colorimetric method with hypochlorite and phenol, l = 630 nm; spectrophotometer Colorimetric determination with sulphonilamide and N-(1-naphtyl) ethylenediamine, l = 520 nm; spectrophotometer Reduction of NO3 (Cu-Cd) with subsequent NO2 determination Oxidation with K2S2O8 with subsequent NO3 determination Visual colorimerty by comparison with artificial standard (Pt-Co) Oxidation with K2Cr2O7 with subsequent determination of K2Cr2O7 excess by titration with FeSO4 (NH4)2SO46H2O Determination as a difference of dissolved oxygen content at the beginning of the experiment and after 5 days incubation at 20 C (bottle method) Gravimetric determination after filtration through 0.45 mm membrane IR spectrometric determination after extraction with CCl4 and Al2O3-column chromatography, l = 2700–3100 cm1; IR-spectrophotometer
6.89–9.57 (8.07) 0.34–15.42 (6.65) 0.003–2.492 (0.190)
Total phosphorus, mg P L1 Electric conductivity, mSm cm1 Carbon dioxide, mg L1 Ammonium, mg N L1 Nitrite, mg N L1 Nitrate, mg N L1 Total nitrogen, mg N L1 Color, degrees Pt-Co Chemical oxygen demand, mg O L1 Biochemical oxygen demand, mg O2 L1 Suspended solids, mg L1 Oil hydrocarbons, mg L1
0.023–2.787 (0.274) 139–1334 (670) 0.0–48.4 (8.7) 0.0–6.271 (0.987) 0.0–0.090 (0.006)
Median 7.92 7.12 0.30 0.097 534 5.7 0.030 0.0
0.006–6.367 (0.031) 0.39–6.44 (1.50) 17–158 (42.3) 16.41–117.56 (55.10)
0.01 0.615 36 50.62
1.03–16.75 (6.11)
4.89
1.2–100.7 (13.11) 0.043–0.215 (0.110)
7.1 0.103
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
maximization between the score matrices of reference and independent data. PLS is widely applied in all fields of analytical chemistry and described in literature in details, see e.g. [24]. Partial robust M-regression (PRM) is a robust multivariate regression technique which employs M-estimator instead of least squares estimator when constructing the regression and the important thing about M-estimator is that it uses weights to take into account the “leverage” of the samples. The results of PRM are usually much more stable against any outlying points compared to PLS. Mathematical introduction of PRM can be found in [25]. PCA and PLS were performed in The Unscrambler1 9.7 (CAMO Software AS, Norway). Model validation procedures are described below in the text for each particular case. CCA and PRM were performed using R 2.15.1 statistical computation software [26]. For PRM calculations we employed ‘‘chemometrics” R package by P. Filzmozer and K. Varmuza [27]. As a measure of predictive ability of the quantitative regression models we used the root mean square error of prediction/calibration
[(Fig._1)TD$IG]
67
(RMSEP and RMSEC correspondingly), which is calculated as: RMSE ¼
sqrtðsumððyi;ref yi;mod Þ2 ; n
where n is the number of samples in the test set, yi,ref is the reference value of the parameter of interest, yi,mod is the value predicted by model for RMSEP or the value initially fitted by the model for RMSEC. 3. Results and discussion PCA was run separately on the data from hydrochemical analysis and ET measurements. The first six score vectors accounted for 97% of variance in ET data and the first seven explained 96% of variance for hydrochemical data. Corresponding score matrices were then subjected to CCA. The extracted values of the three first canonic roots were 0.964, 0.851, 0.547, which implies that the reasonable match of the
Fig. 1. CCA similarity maps defined by canonical variates 1 and 2, for (a) ET analysis, and (b) hydrochemical measurements in pond waters.
68
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
Table 2 Parameters of ET performance for prediction of total nitrogen, COD and conductivity in pond water samples. R2
Slope
Offset
0.82 0.86
0.22 0.05
0.82 0.81
0.71 1.11
Chemical oxygen demand (16.41–117.56), mg O L1 Calibration 0.91 Validation 1.03
5.07 -4.74
0.91 0.79
9.23 11.08
Conductivity (139–1334), mSm cm1 Calibration Validation
82 78
0.88 0.91
RMSE
1
Total nitrogen (0.39–6.44), mg N L Calibration Validation
0.88 0.84
variance structure in two data sets was observed. In CCA similarity maps (Fig. 1) one can compare e.g. the location of samples #53, 55 and 17 in the lower part of the plots and #50, 41, 42 in the upper part. This result assumes that there is a possibility of developing certain regression models for numerical prediction of particular hydrochemical parameters from ET data. PLS models for estimation of CODCr, total nitrogen and conductivity were constructed as examples of such correlations. Table 2 shows the parameters of “measured vs predicted” lines for these models. Validation was performed with independent test set obtained from random split of samples into calibration (41 samples) and validation (17 samples). A reasonably good prediction performance can be observed for these three parameters. Of particular practical interest is a good correlation of the ET data with CODCr parameter since the traditional procedure for its determination is quite long and tedious (sample oxidation with potassium bichromate and titration of its excess). A correlation with conductivity is not surprising since potentiometric response of the sensors is due to the presence of the ions, which are also responsible for conductive properties of the sample. The rest of the hydrochemical parameters, however, were not that well correlated with ET response. Typical R2 values in corresponding PLS regression models were around 0.6. This is quite well understood taking into account very complex composition of the surface water samples. Such parameters as e.g. total and inorganic phosphorus and carbon dioxide can hardly be correctly measured by potentiometric sensors, due to a high hydrophilicity of the corresponding anions (phosphate and carbonate). An attempt was made to predict the toxicity values from hydrochemical data by constructing a PLS1 regression model with Daphnia bioassay data as an independent variable. The parameters of model validation were very poor (R2 = 0.28, RMSEP = 30% in full cross-validation), which means that hydrochemical parameters are poorly connected to Daphnia response, although they can still provide valuable information on water quality. Hotelling’s T2 test for two-dimensional score plot of samples revealed several outliers, however, their subsequent elimination did not improve the parameters of the PLS regression. As the next step in this study we constructed PLS regression model for prediction of toxicity in terms of Daphnia response from ET measurements. The first attempt resulted in very poor correlation (R2 = 0.25). A thorough inspection of X–Y relation outliers plot detected several possible outliers, namely the samples 4, 8, 27, 28 and 29 both from the surface and the bottom. After elimination of these samples from the model R2 of 0.85 and RMSEP of 14% were obtained in full cross-validation. This confirms our previous findings in [17] that bioassay results can be simulated with reasonable precision by a multisensor system. The reason for which the above mentioned samples did not fit the model was
137 128
explored on the basis of hydrochemical data, corresponding pond location and other individual features. It was found that pond 4 had very strong organic pollution confirmed by high values of total microbial count from additional microbiological analysis. It appears that ET system exhibits certain sensitivity to these unknown organic substances but Daphnia do not have it (corresponding toxicity values were moderate: 36.7% for surface and 40% for near-bottom layers) and, consequently, 4s and 4b samples became outliers. The pond 8 is located along the road with heavy traffic and was possibly contaminated with heavy metals and hydrocarbons. Another issue regarding this pond is that it had significant amount of cyanobacteria, although it is not yet clearly understood now how it can be related to ET response. The ponds 27 and 29 were supposedly polluted by organic substances, confirmed by microbial analysis, just like in case with the pond 8. The nature of deviations of pond 28’s water remained unclear. In the presence of outliers robust regression methods can perform better [28], thus we attempted to use PRM regression instead of PLS to see if it can deal with outlying samples without the need to eliminate them from the model. The resulted model with all of the available samples also exhibited poor correlation (R2 = 0.55 in cross-validation) although higher than that of ordinary PLS (see above). Subsequent removal of the data from the ponds 4, 8, 27, 28, 29 resulted in a reasonable PRM regression model with R2 = 0.83. It is widely known that cross-validation of the multivariate regression models can often produce over-optimistic results [29– 31]. To avoid any far-reaching conclusions we performed validation of regression models for toxicity prediction using an independent test set. This was done both for PLS and PRM models. The following procedure was employed for splitting the samples into calibration and test set: for each of the analyzed ponds one sample from the surface or from the near-bottom layer was randomly chosen for calibration set while the remaining one for each pond was employed in the test set to assess the predictive ability of the ET system. The results of predictions are given in the Fig. 2. It can be seen that both methods allow for reasonable precision in toxicity assessment from potentiometric ET data. Summarized RMSEP values for these test sets were 18% for PLS and 20% for PRM regression. Noticeable deviations from the reference toxicity values can be observed for samples 18s and 24b in both cases. Taking into account numerous advantages of ET methodology such as fast and simple measurement protocol, absence of specific reagents, this technique can be suggested as a potential alternative for traditional bioassay methods. Due to sufficient number of samples in this study we were able to validate the regression models in a reliable and statistically significant way, thus the applicability of potentiometric multisensor systems for Daphnia magna bioassay simulation was reliably proved. What is particularly interesting about multisensor approach is that toxicity
[(Fig._2)TD$IG]
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
69
Fig. 2. Prediction of toxicity in test set samples with PLS (a) and PRM (b) regression models.
predictions in terms of various bioassays are potentially possible with the same sensor system. The conclusion can be done from a single measurement by a sensor system, using preliminary obtained calibration models for each bioassay type. Obviously, sensitivity of the sensor array in this case should cover the whole toxicants range, which the reference bioassays are sensitive to. It is worth of mentioning that the regulatory limits established for particular pollutants in potable water can be significantly lower than sensitivity thresholds both for daphnia and potentiometric sensors, thus proper care should be taken when translating these studies into neighboring fields. Another important issue is that we made the calibrations against the 96-h daphnia tests and the results of these tests do not take in account contaminants with prolonged and delayed toxicity. This gives a chance for underestimation of potential risks related to some herbicides, polyaromatic compounds, etc. To take this into account a calibration against prolonged toxicity tests would be required. We see the primary tasks for further development in the improvement of precision of this bioassay simulation and in clarification of platform applicability limits by examination of other bioassay methods as the source of reference data for the multisensor array.
4. Conclusion ET methodology adopted for the toxicity evaluation in terms of bioassay opens up a perspective of routine water quality and toxicity evaluation by such systems without using biological objects at all. A multisensor system, being calibrated against the response of these various biological objects, can be applied for assessing the potential harm of contaminated water to these objects. Toxicity values of urban water samples in this study were determined by potentiometric electronic tongue in terms of Daphnia magna death rate. Predictive models constructed from ET data (PLS and PRM regression) were validated with independent test sets. It was found that root mean square prediction errors did not exceed 20%. We believe this is a very promising result since the analytical task formulation is quite complex and rather unusual. A multisensor approach can offer advantages of on-line measurements and continuous monitoring, simple and inexpensive experimental set-up and reagentless procedure. However, further extensive research in this field is required to increase the precision of the method and to understand the reasons for certain observed outliers.
70
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
Acknowledgement This work was partially financially supported by Government of Russian Federation, Grant 074-U01. References [1] EU Directive, 200. Directive 2000/60/EC of the European Parliament and of the Council of 23 October 2000 establishing a framework for Community action in the field of water policy. Official Journal L 327, 22/12/2000 P. 0001–0073. http://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=CELEX:32000L0060: EN:HTML, last accessed on 12 October 2013. [2] US Senate, 2002. Federal Water Pollution Control Act, http://www.epw.senate. gov/water.pdf, last accessed on 12 October 2013. [3] C.J. Keddy, J.C. Greene, M.A. Bonnell, Ecotoxicology and Environmental Safety 30 (1995) 221. [4] C. Sonnenschein, A.M. Soto, Journal of Steroid Biochemistry 65 (1998) 143. [5] ISO 11348-1, 2007. Water quality – Determination of the inhibitory effect of water samples on the light emission of Vibrio fischeri(Luminescent bacteria test) – Part 1: Method using freshly prepared bacteria. [6] ISO 6341, 2012. Water quality – Determination of the inhibition of the mobility of Daphnia magna Straus (Cladocera, Crustacea) – Acute toxicity test. [7] ISO 8692, 2012. Water quality – Fresh water algal growth inhibition test with unicellular green algae. [8] R.S.J. Alkasir, M. Ornatska, S. Andreescu, Analytical Chemistry 84 (2012) 9729. [9] J. Li, Y. Yu, Y. Wang, J. Qian, J. Zhi, C. Blaise, J.-F. Férard (Eds.), Electrochimica Acta (2013). [10] B. Thomas Johnson, Microtox1 Acute Toxicity Test, in: C. Blaise, J.-F. Férard (Eds.), Small-scale Freshwater Toxicity Investigations, vol. 1 Toxicity Test Methods, Springer, Berlin, 2005 pp. 69–105. [11] P. Ciosek, W. Wroblewski, Talanta 76 (2008) 548. [12] Z. Wei, J. Wang, Analytica Chimica Acta 694 (2011) 45.
[13] X. Cetó, J.M. Gutiérrez, M. Gutiérrez, F. Céspedes, J. Capdevila, S. Mínguez, C. Jiménez-Jorquera, M. Del Valle, Analytica Chimica Acta 732 (2012) 172. [14] A. Legin, A. Rudnitskaya, Yu. Vlasov, C. Di Natale, E. Mazzone, A. D‘Amico, Sensors and Actuators B: Chemical 65 (2000) 232. [15] A. Riul Jr., C.A.R. Dantas, C.M. Miyazaki, O.N. Oliveira Jr., Analyst 135 (2010) 2481. [16] H. Smyth, D. Cozzolino, Chemical Reviews 113 (2013) 1429. [17] D. Kirsanov, O. Zadorozhnaya, A. Krasheninnikov, N. Komarova, A. Popov, A. Legin, Sensors and Actuators B: Chemical 179 (2013) 282. [18] D. Kirsanov, M. Khaydukova, L. Tkachenko, A. Legin, V. Babain, Electroanalysis 24 (2012) 121–130. [19] P. Ciosek, W. Wroblewski, Talanta 76 (2008) 548–556. [20] G. Yu. Mourzina, M.J. Schöning, J. Schubert, W. Zander, A.V. Legin, G. Yu. Vlasov, H. Lüth, Analytica Chimica Acta 433 (2001) 103–110. [21] K.H. Esbensen, Multivariate data analysis – in practice, An Introduction to Multivariate Data Analysis and Experimental Design, 5th ed., CAMO AS Publ., Oslo, 2001. [22] J.F. Hair, R.E. Anderson, R.L. Tatham, W.C. Black, Multivariate Data Analysis, 5th ed., Prentice Hall Inc., USA, 1998. [23] D. Kirsanov, O. Mednova, V. Vietoris, P.A. Kilmartin, A. Legin, Talanta 90 (2012) 109. [24] S. Wold, M. Sjöström, L. Eriksson, Chemometrics and Intelligent Laboratory Systems 58 (2001) 109. [25] S. Serneels, C. Croux, P. Filzmoser, P.J. Van Espen, Chemometrics and Intelligent Laboratory Systems 79 (2005) 55. [26] R Development Core Team, R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria, 2010http://www.R-project.orghttp://cran.r-project.org/web/packages/chemometrics/index.html. [27] http://cran.r-project.org/web/packages/chemometrics/index.html [28] M. Daszykowski, K. Kaczmarek, Y. Vander Heyden, B. Walczak, Chemometrics and Intelligent Laboratory Systems 85 (2007) 203. [29] R.G. Brereton, TrAC-Trends in Analytical Chemistry 25 (2006) 1103. [30] P. Filzmoser, B. Liebmann, K. Varmuza, Journal of Chemometrics 33 (2009) 160. [31] K.H. Esbensen, P. Geladi, Journal of Chemometrics 24 (2010) 168.
Contents lists available at ScienceDirect
Analytica Chimica Acta journal homepage: www.elsevier.com/locate/aca
Mimicking Daphnia magna bioassay performance by an electronic tongue for urban water quality control Dmitry Kirsanov a,b, *, Evgeny Legin b,c , Anatoly Zagrebin d , Natalia Ignatieva d, Vladimir Rybakin d , Andrey Legin a,b a
Laboratory of Chemical Sensors, St. Petersburg State University, St. Petersburg, Russia Laboratory of Artificial Sensor Systems, ITMO University, St. Petersburg, Russia Sensor Systems LLC, St. Petersburg, Russia d Institute of Limnology, Russian Academy of Sciences, St. Petersburg, Russia b c
H I G H L I G H T S
G R A P H I C A L A B S T R A C T
Daphnia magna bioassay can be simulated with multisensor system. Urban water toxicity can be predicted from potentiometric ET data. Independent test set validation confirms statistical significance of the results.
A R T I C L E I N F O
A B S T R A C T
Article history: Received 13 November 2013 Received in revised form 14 March 2014 Accepted 15 March 2014 Available online 20 March 2014
Toxicity is one of the key parameters of water quality in environmental monitoring. However, being evaluated as a response of living beings (as their mobility, fertility, death rate, etc.) to water quality, toxicity can only be assessed with the help of these living beings. This imposes certain restrictions on toxicity bioassay as an analytical method: biotest organisms must be properly bred, fed and kept under strictly regulated conditions and duration of tests can be quite long (up to several days), thus making the whole procedure the prerogative of the limited number of highly specialized laboratories. This report describes an original application of potentiometric multisensor system (electronic tongue) when the set of electrochemical sensors was calibrated against Daphnia magna death rate in order to perform toxicity assessment of urban waters without immediate involvement of living creatures. PRM (partial robust M) and PLS (projections on latent structures) regression models based on the data from this multisensor system allowed for prediction of toxicity of unknown water samples in terms of biotests but in the fast and simple instrumental way. Typical errors of water toxicity predictions were below 20% in terms of Daphnia death rate which can be considered as a good result taking into account the complexity of the task. ã 2014 Elsevier B.V. All rights reserved.
Keywords: Bioassay Water toxicity Multisensor systems Electronic tongue
1. Introduction Water quality and safety is a question of highest importance nowadays. Extensive anthropogenic influence led to massive
* Corresponding author. Tel.: +7 812 328 28 35. E-mail address: [email protected] (D. Kirsanov). http://dx.doi.org/10.1016/j.aca.2014.03.021 0003-2670/ ã 2014 Elsevier B.V. All rights reserved.
hydrosphere pollution all over the world. The concerns of the modern society regarding this issue are reflected in numerous legislative initiatives in this field, such as European Union Water Framework Directive [1], Clean Water Act [2]. The number of analytical techniques and methods suggested for water quality control is really huge and growing. It would be hard to name an analytical method which was not yet suggested for water analysis. Chromatographic, electrochemical, optical, etc. methods can be
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
effectively employed to solve certain particular tasks in this field. One of the integral parameters which are of ultimate importance in ecological water control is its toxicity. Toxicity is understood as a measure of a real harm to the living creatures caused by water sample. There is a special group of methods called bioassay which is intended for direct estimation of water toxicity for biological objects. The general idea behind this type of analysis is to monitor a biological reaction of a biotest object placed in a water sample under study. This reaction is measured in some appropriate way and compared with that of the same biotest object placed in a control sample of pure nontoxic water. Growth rate, mobility, survivorship rate of the test organisms, etc. can be monitored as biological reactions [3]. Various aquatic biota species can be employed for testing, such as e.g. algae, luminescent bacteria, infusoria, daphnia, fish, etc. An obvious advantage of this approach is direct information about biological harm that can be caused to the test object by contaminated water, while most of the other methods based on physicochemical measurements can only provide information on the content of selected chemical substances. This is, of course, useful but not always directly related to the danger posed for living beings, especially when several contaminants are present in the water simultaneously that is quite common. It is important mentioning that sometimes different substances with dissimilar chemical structures can cause the same degree of biological damage in various organisms [4]. Different bioassays and test procedures were suggested and some of them are now accepted as standard analysis legitimized in ISO standards, e.g. Vibrio fischeri (luminescent bacteria) [5], Daphnia magna [6], green algae [7]. It is noteworthy that none of the suggested methods can serve as a universal toxicity assessment instrument because of the various sensitivity patterns of different bioassays. Being an indispensable technique for toxicity assessment bioassay methodology still has a number of drawbacks: its implementation requires highly specialized laboratories; most of the methods are quite slow (usually it takes 24–72 h to get the results); biotest objects must be properly fed and kept under thoroughly regulated conditions. There are numerous publications in literature where attempts were made to overcome these bioassay drawbacks by developing new sensor technologies, see e.g. recent reports [8,9]. The U.S. Environmental Protection Agency (EPA) has certified several Rapid Toxicity Testing Systems (RTTS), like e.g. MicroTox1 [10]. A thorough inspection of the protocols of such methods reveals that all of them require either some biological objects (e.g. freeze-dried bacteria, usually luminescent Vibrio fischeri) or bio/ chemical reagents or quite sophisticated preparative procedures. Besides, some of these methods require highly qualified personnel for operation and in many cases reagent sets include bacteria or living cells, which makes them costly and imply special conditions for their storage. This makes such RTTS and the like methods hardly suitable for continuous autonomous monitoring of water toxicity. Electronic tongue (ET) methodology proved to be useful in complex assessment of various liquid media [11–13]. The operation principle of such systems is quite simple – it suggests the use of an array of cross-sensitive chemical sensors and subsequent processing of unresolved analytical signal from a sensor array by modern multivariate statistic methods [14]. A cross-sensitivity of sensors means that each sensor of the array has its own sensitivity and selectivity pattern and can respond with different response value to various substances of interest in a sample. The resulted signal from all sensors of such array contains information on overall sample composition (according to sensitivity patterns of the sensors) and can be related to certain quality parameters of the samples by means of multivariate data processing both in qualitative and quantitative way. The ET research field is rapidly
65
growing due to a distinct trend in modern analytical chemistry towards fast and simple instruments [15,16]. Our preliminary research [17] with various aquatic organisms (Daphnia magna, Chlorella vulgaris, Paramecium caudatum) has shown that potentiometric multisensor system being preliminary calibrated against bioassay results for a set of samples with known toxicity allows for prediction of toxicity in new, totally unknown water samples in terms of bioassay (e.g. survivorship rate). This procedure does not require complex analytical instrumentation, sophisticated sample preparation and measuring procedures and elaborated biological object maintenance. The idea of the present study was to extend our preliminary studies by a full-scale experiment with sufficient number of samples to explore the opportunities and limitations of ET technology for water toxicity detection. For this purpose we participated in ongoing research initiative devoted to the continuous monitoring of urban water reservoirs in St. Petersburg (Russia). In the framework of this activity regular hydrochemical and toxicity analysis of pond water had been performed, thus we were able to analyze sufficient amount of samples by potentiometric multisensor system. A full access to the results of hydrochemical analysis and toxicity assessment permitted proper statistical validation of the ET predictive performance.
2. Materials and methods 2.1. Samples A series of pond water samples was collected during several experimental sessions in July and August 2012. 29 ponds located in St. Petersburg city area and its closest suburbs were sampled. Water samples from each pond were taken from two horizons – surface layer (0.5 m thick) and near-bottom layer, yielding 58 samples in total. The two samples from different horizons of the same pond were not mixed together and were treated separately in all further studies. Each sample was split into three portions: the first one for classical hydrochemical analysis, the second one for bioassay with Daphnia magna and the third one for electronic tongue measurements. All samples had the same history of storage and transportation before various analyses. Due to diverse location of sampling various pond water samples had various visual appearance, some of them were clear and transparent while the others were turbid, with brownish color and distinct smell. All samples were encoded by a number and a letter: thus 15s stays for the sample taken from the surface (s) of the 15th pond and e.g. 3b stays for the sample taken from the bottom (b) of the 3rd pond. The depth values for the studied ponds varied in the range 1–4.8 m with the average value 2.1 m; traffic conditions in the immediate proximity of the ponds were from heavy to zero (for several ponds located in the park areas). Temperature of the sampled water was in the range 11–24 C with the average value 18.7 C.
2.2. Hydrochemical analysis 15 different hydrochemical parameters were determined for all samples, these were pH, dissolved oxygen, total phosphorus (TP), inorganic phosphorus (IP), electric conductivity, carbon dioxide, ammonium, nitrite, nitrate, total nitrogen, color, chemical oxygen demand (CODCr), biochemical oxygen demand, suspended solids, oil hydrocarbons. The methods of their assessment and their range in the studied samples are presented in Table 1. The determination of these parameters is regulated by normative documents developed by Russian Federal State Budgetary Institution “State Hydrochemical Institute”.
66
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
2.3. Daphnia magna bioassay One- or two-day-old Daphnia magna Straus were employed in 96 h toxicity tests. Ten animals were used for a study in each sample and all studies were performed in three replicate runs. 25 mL of sample was used for each replica. The sample was considered being toxic when 50% or more of the animals died. The fact of death was determined visually using optical microscope. The data from three replicas were averaged for further processing. Resulted toxicity values varied from 0 to 100% with the average 44.4% and the median 35.0%. Due to technical issues bioassay results were obtained only for 50 samples. 2.4. Multisensor system analysis Multisensor system (Electronic Tongue) employed in this study was based on 19 potentiometric cross-sensitive sensors. Our previous trial [17] employed 23 sensors, however during the data processing we found that only 19 sensors have reasonably high regression coefficients in PLS models for prediction of water toxicity in terms of Daphnia magna bioassay while other were useful for infusoria and chlorella prediction. Thus, only these 19 sensors were employed for described experiment. Six of them were poly(vinylchloride) (PVC)-plasticized anion-sensitive sensors based on various anion-exchangers, seven were PVC-plasticized cation-sensitive sensitive sensors with neutral ligands described in Ref. [18], and six sensors were chalcogenide glass electrodes with various RedOx sensitivity patterns. The details on sensor compositions and preparation procedures are widely available in literature, see e.g. [18–20]. Electrochemical measurements were carried out in the following galvanic cell: Cu|Ag|AgCl, KClsat|sample solution|sensor membrane|solid contact|Cu EMF values (sensor potentials) were measured with 0.1 mV precision against the standard silver/silver chloride reference electrode using made-in-house 32-channel digital high impedance voltmeter connected to a PC for data acquisition. All samples were measured in four physically different replicas and the results were averaged for further processing. Replicated measurements of different samples were taken in random order. Water samples were measured as is, without any sample preparation procedures,
dilution, etc. The measurement time in each sample was 3 min and three last readings were averaged and further employed for data processing. After the measurements sensors were washed by several portions of distilled water for 2–3 min. This procedure allowed for 5 mV reproducibility in replicated measurements of the same samples. The measurements in pond water samples with multisensor system were performed in parallel with bioassay measurements in physically different portions of the same pond waters. 2.5. Data processing The results of potentiometric measurements in the samples were combined into matrix form with samples in rows and sensors in columns. Thus, each matrix element was the reading of the sensor k in the sample i. The dimensions of the ET matrix were 58 samples 19 sensors. The same was done for the data of hydrochemical measurements, which yielded the matrix 58 samples 15 parameters. The data from bioassay resulted in a vector 50 samples 1 parameter. We applied several different techniques to relate these two data sets with each other and also to relate multisensor system response with toxicity values produced by Daphnia magna bioassay. Principal component analysis (PCA), canonical correlation analysis (CCA), partial least squares regression (PLS) and partial robust M-regression (PRM) were employed for data analysis. PCA is a widely applied method for data dimensionality reduction and visualization of hidden data structure. The method is based on projections of initial samples from initial multivariate space onto a new coordinate space where new coordinate axes (principal components) are located in the direction of maximum variance and are mutually orthogonal; see e.g. [21] for details. CCA is aimed for the assessment of the correlation between two data sets sharing the same row mode i.e. obtained for the same sample set [22]. To apply CCA we preliminary compressed the ET and hydrochemical data with PCA without significant information loss and after that CCA was run on the resulting PC scores for both data sets. This approach was shown to be effective for ET data processing in wine analysis [23]. PLS is a classical tool for multivariate regression, it finds the regression coefficients vector with a condition of covariance
Table 1 Summary on hydrochemical methods. Parameter
Analytical method and instrumentation
Range min–max (average)
pH, pH units Dissolved oxygen, mg L1 Inorganic phosphorus, mg P L1
Electrometric determination by glass electrode, pH-meter (25 C) Titrimetric determination by Winkler (iodometric method) Colorimetric determination with ammonium molybdate and ascorbic acid, l = 880 nm; spectrophotometer Oxidation with K2S2O8 with subsequent IP determination Conductometric determination in situ (25 C); conductometer Calculative method based on pH and HCO3 data Colorimetric method with hypochlorite and phenol, l = 630 nm; spectrophotometer Colorimetric determination with sulphonilamide and N-(1-naphtyl) ethylenediamine, l = 520 nm; spectrophotometer Reduction of NO3 (Cu-Cd) with subsequent NO2 determination Oxidation with K2S2O8 with subsequent NO3 determination Visual colorimerty by comparison with artificial standard (Pt-Co) Oxidation with K2Cr2O7 with subsequent determination of K2Cr2O7 excess by titration with FeSO4 (NH4)2SO46H2O Determination as a difference of dissolved oxygen content at the beginning of the experiment and after 5 days incubation at 20 C (bottle method) Gravimetric determination after filtration through 0.45 mm membrane IR spectrometric determination after extraction with CCl4 and Al2O3-column chromatography, l = 2700–3100 cm1; IR-spectrophotometer
6.89–9.57 (8.07) 0.34–15.42 (6.65) 0.003–2.492 (0.190)
Total phosphorus, mg P L1 Electric conductivity, mSm cm1 Carbon dioxide, mg L1 Ammonium, mg N L1 Nitrite, mg N L1 Nitrate, mg N L1 Total nitrogen, mg N L1 Color, degrees Pt-Co Chemical oxygen demand, mg O L1 Biochemical oxygen demand, mg O2 L1 Suspended solids, mg L1 Oil hydrocarbons, mg L1
0.023–2.787 (0.274) 139–1334 (670) 0.0–48.4 (8.7) 0.0–6.271 (0.987) 0.0–0.090 (0.006)
Median 7.92 7.12 0.30 0.097 534 5.7 0.030 0.0
0.006–6.367 (0.031) 0.39–6.44 (1.50) 17–158 (42.3) 16.41–117.56 (55.10)
0.01 0.615 36 50.62
1.03–16.75 (6.11)
4.89
1.2–100.7 (13.11) 0.043–0.215 (0.110)
7.1 0.103
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
maximization between the score matrices of reference and independent data. PLS is widely applied in all fields of analytical chemistry and described in literature in details, see e.g. [24]. Partial robust M-regression (PRM) is a robust multivariate regression technique which employs M-estimator instead of least squares estimator when constructing the regression and the important thing about M-estimator is that it uses weights to take into account the “leverage” of the samples. The results of PRM are usually much more stable against any outlying points compared to PLS. Mathematical introduction of PRM can be found in [25]. PCA and PLS were performed in The Unscrambler1 9.7 (CAMO Software AS, Norway). Model validation procedures are described below in the text for each particular case. CCA and PRM were performed using R 2.15.1 statistical computation software [26]. For PRM calculations we employed ‘‘chemometrics” R package by P. Filzmozer and K. Varmuza [27]. As a measure of predictive ability of the quantitative regression models we used the root mean square error of prediction/calibration
[(Fig._1)TD$IG]
67
(RMSEP and RMSEC correspondingly), which is calculated as: RMSE ¼
sqrtðsumððyi;ref yi;mod Þ2 ; n
where n is the number of samples in the test set, yi,ref is the reference value of the parameter of interest, yi,mod is the value predicted by model for RMSEP or the value initially fitted by the model for RMSEC. 3. Results and discussion PCA was run separately on the data from hydrochemical analysis and ET measurements. The first six score vectors accounted for 97% of variance in ET data and the first seven explained 96% of variance for hydrochemical data. Corresponding score matrices were then subjected to CCA. The extracted values of the three first canonic roots were 0.964, 0.851, 0.547, which implies that the reasonable match of the
Fig. 1. CCA similarity maps defined by canonical variates 1 and 2, for (a) ET analysis, and (b) hydrochemical measurements in pond waters.
68
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
Table 2 Parameters of ET performance for prediction of total nitrogen, COD and conductivity in pond water samples. R2
Slope
Offset
0.82 0.86
0.22 0.05
0.82 0.81
0.71 1.11
Chemical oxygen demand (16.41–117.56), mg O L1 Calibration 0.91 Validation 1.03
5.07 -4.74
0.91 0.79
9.23 11.08
Conductivity (139–1334), mSm cm1 Calibration Validation
82 78
0.88 0.91
RMSE
1
Total nitrogen (0.39–6.44), mg N L Calibration Validation
0.88 0.84
variance structure in two data sets was observed. In CCA similarity maps (Fig. 1) one can compare e.g. the location of samples #53, 55 and 17 in the lower part of the plots and #50, 41, 42 in the upper part. This result assumes that there is a possibility of developing certain regression models for numerical prediction of particular hydrochemical parameters from ET data. PLS models for estimation of CODCr, total nitrogen and conductivity were constructed as examples of such correlations. Table 2 shows the parameters of “measured vs predicted” lines for these models. Validation was performed with independent test set obtained from random split of samples into calibration (41 samples) and validation (17 samples). A reasonably good prediction performance can be observed for these three parameters. Of particular practical interest is a good correlation of the ET data with CODCr parameter since the traditional procedure for its determination is quite long and tedious (sample oxidation with potassium bichromate and titration of its excess). A correlation with conductivity is not surprising since potentiometric response of the sensors is due to the presence of the ions, which are also responsible for conductive properties of the sample. The rest of the hydrochemical parameters, however, were not that well correlated with ET response. Typical R2 values in corresponding PLS regression models were around 0.6. This is quite well understood taking into account very complex composition of the surface water samples. Such parameters as e.g. total and inorganic phosphorus and carbon dioxide can hardly be correctly measured by potentiometric sensors, due to a high hydrophilicity of the corresponding anions (phosphate and carbonate). An attempt was made to predict the toxicity values from hydrochemical data by constructing a PLS1 regression model with Daphnia bioassay data as an independent variable. The parameters of model validation were very poor (R2 = 0.28, RMSEP = 30% in full cross-validation), which means that hydrochemical parameters are poorly connected to Daphnia response, although they can still provide valuable information on water quality. Hotelling’s T2 test for two-dimensional score plot of samples revealed several outliers, however, their subsequent elimination did not improve the parameters of the PLS regression. As the next step in this study we constructed PLS regression model for prediction of toxicity in terms of Daphnia response from ET measurements. The first attempt resulted in very poor correlation (R2 = 0.25). A thorough inspection of X–Y relation outliers plot detected several possible outliers, namely the samples 4, 8, 27, 28 and 29 both from the surface and the bottom. After elimination of these samples from the model R2 of 0.85 and RMSEP of 14% were obtained in full cross-validation. This confirms our previous findings in [17] that bioassay results can be simulated with reasonable precision by a multisensor system. The reason for which the above mentioned samples did not fit the model was
137 128
explored on the basis of hydrochemical data, corresponding pond location and other individual features. It was found that pond 4 had very strong organic pollution confirmed by high values of total microbial count from additional microbiological analysis. It appears that ET system exhibits certain sensitivity to these unknown organic substances but Daphnia do not have it (corresponding toxicity values were moderate: 36.7% for surface and 40% for near-bottom layers) and, consequently, 4s and 4b samples became outliers. The pond 8 is located along the road with heavy traffic and was possibly contaminated with heavy metals and hydrocarbons. Another issue regarding this pond is that it had significant amount of cyanobacteria, although it is not yet clearly understood now how it can be related to ET response. The ponds 27 and 29 were supposedly polluted by organic substances, confirmed by microbial analysis, just like in case with the pond 8. The nature of deviations of pond 28’s water remained unclear. In the presence of outliers robust regression methods can perform better [28], thus we attempted to use PRM regression instead of PLS to see if it can deal with outlying samples without the need to eliminate them from the model. The resulted model with all of the available samples also exhibited poor correlation (R2 = 0.55 in cross-validation) although higher than that of ordinary PLS (see above). Subsequent removal of the data from the ponds 4, 8, 27, 28, 29 resulted in a reasonable PRM regression model with R2 = 0.83. It is widely known that cross-validation of the multivariate regression models can often produce over-optimistic results [29– 31]. To avoid any far-reaching conclusions we performed validation of regression models for toxicity prediction using an independent test set. This was done both for PLS and PRM models. The following procedure was employed for splitting the samples into calibration and test set: for each of the analyzed ponds one sample from the surface or from the near-bottom layer was randomly chosen for calibration set while the remaining one for each pond was employed in the test set to assess the predictive ability of the ET system. The results of predictions are given in the Fig. 2. It can be seen that both methods allow for reasonable precision in toxicity assessment from potentiometric ET data. Summarized RMSEP values for these test sets were 18% for PLS and 20% for PRM regression. Noticeable deviations from the reference toxicity values can be observed for samples 18s and 24b in both cases. Taking into account numerous advantages of ET methodology such as fast and simple measurement protocol, absence of specific reagents, this technique can be suggested as a potential alternative for traditional bioassay methods. Due to sufficient number of samples in this study we were able to validate the regression models in a reliable and statistically significant way, thus the applicability of potentiometric multisensor systems for Daphnia magna bioassay simulation was reliably proved. What is particularly interesting about multisensor approach is that toxicity
[(Fig._2)TD$IG]
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
69
Fig. 2. Prediction of toxicity in test set samples with PLS (a) and PRM (b) regression models.
predictions in terms of various bioassays are potentially possible with the same sensor system. The conclusion can be done from a single measurement by a sensor system, using preliminary obtained calibration models for each bioassay type. Obviously, sensitivity of the sensor array in this case should cover the whole toxicants range, which the reference bioassays are sensitive to. It is worth of mentioning that the regulatory limits established for particular pollutants in potable water can be significantly lower than sensitivity thresholds both for daphnia and potentiometric sensors, thus proper care should be taken when translating these studies into neighboring fields. Another important issue is that we made the calibrations against the 96-h daphnia tests and the results of these tests do not take in account contaminants with prolonged and delayed toxicity. This gives a chance for underestimation of potential risks related to some herbicides, polyaromatic compounds, etc. To take this into account a calibration against prolonged toxicity tests would be required. We see the primary tasks for further development in the improvement of precision of this bioassay simulation and in clarification of platform applicability limits by examination of other bioassay methods as the source of reference data for the multisensor array.
4. Conclusion ET methodology adopted for the toxicity evaluation in terms of bioassay opens up a perspective of routine water quality and toxicity evaluation by such systems without using biological objects at all. A multisensor system, being calibrated against the response of these various biological objects, can be applied for assessing the potential harm of contaminated water to these objects. Toxicity values of urban water samples in this study were determined by potentiometric electronic tongue in terms of Daphnia magna death rate. Predictive models constructed from ET data (PLS and PRM regression) were validated with independent test sets. It was found that root mean square prediction errors did not exceed 20%. We believe this is a very promising result since the analytical task formulation is quite complex and rather unusual. A multisensor approach can offer advantages of on-line measurements and continuous monitoring, simple and inexpensive experimental set-up and reagentless procedure. However, further extensive research in this field is required to increase the precision of the method and to understand the reasons for certain observed outliers.
70
D. Kirsanov et al. / Analytica Chimica Acta 824 (2014) 64–70
Acknowledgement This work was partially financially supported by Government of Russian Federation, Grant 074-U01. References [1] EU Directive, 200. Directive 2000/60/EC of the European Parliament and of the Council of 23 October 2000 establishing a framework for Community action in the field of water policy. Official Journal L 327, 22/12/2000 P. 0001–0073. http://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=CELEX:32000L0060: EN:HTML, last accessed on 12 October 2013. [2] US Senate, 2002. Federal Water Pollution Control Act, http://www.epw.senate. gov/water.pdf, last accessed on 12 October 2013. [3] C.J. Keddy, J.C. Greene, M.A. Bonnell, Ecotoxicology and Environmental Safety 30 (1995) 221. [4] C. Sonnenschein, A.M. Soto, Journal of Steroid Biochemistry 65 (1998) 143. [5] ISO 11348-1, 2007. Water quality – Determination of the inhibitory effect of water samples on the light emission of Vibrio fischeri(Luminescent bacteria test) – Part 1: Method using freshly prepared bacteria. [6] ISO 6341, 2012. Water quality – Determination of the inhibition of the mobility of Daphnia magna Straus (Cladocera, Crustacea) – Acute toxicity test. [7] ISO 8692, 2012. Water quality – Fresh water algal growth inhibition test with unicellular green algae. [8] R.S.J. Alkasir, M. Ornatska, S. Andreescu, Analytical Chemistry 84 (2012) 9729. [9] J. Li, Y. Yu, Y. Wang, J. Qian, J. Zhi, C. Blaise, J.-F. Férard (Eds.), Electrochimica Acta (2013). [10] B. Thomas Johnson, Microtox1 Acute Toxicity Test, in: C. Blaise, J.-F. Férard (Eds.), Small-scale Freshwater Toxicity Investigations, vol. 1 Toxicity Test Methods, Springer, Berlin, 2005 pp. 69–105. [11] P. Ciosek, W. Wroblewski, Talanta 76 (2008) 548. [12] Z. Wei, J. Wang, Analytica Chimica Acta 694 (2011) 45.
[13] X. Cetó, J.M. Gutiérrez, M. Gutiérrez, F. Céspedes, J. Capdevila, S. Mínguez, C. Jiménez-Jorquera, M. Del Valle, Analytica Chimica Acta 732 (2012) 172. [14] A. Legin, A. Rudnitskaya, Yu. Vlasov, C. Di Natale, E. Mazzone, A. D‘Amico, Sensors and Actuators B: Chemical 65 (2000) 232. [15] A. Riul Jr., C.A.R. Dantas, C.M. Miyazaki, O.N. Oliveira Jr., Analyst 135 (2010) 2481. [16] H. Smyth, D. Cozzolino, Chemical Reviews 113 (2013) 1429. [17] D. Kirsanov, O. Zadorozhnaya, A. Krasheninnikov, N. Komarova, A. Popov, A. Legin, Sensors and Actuators B: Chemical 179 (2013) 282. [18] D. Kirsanov, M. Khaydukova, L. Tkachenko, A. Legin, V. Babain, Electroanalysis 24 (2012) 121–130. [19] P. Ciosek, W. Wroblewski, Talanta 76 (2008) 548–556. [20] G. Yu. Mourzina, M.J. Schöning, J. Schubert, W. Zander, A.V. Legin, G. Yu. Vlasov, H. Lüth, Analytica Chimica Acta 433 (2001) 103–110. [21] K.H. Esbensen, Multivariate data analysis – in practice, An Introduction to Multivariate Data Analysis and Experimental Design, 5th ed., CAMO AS Publ., Oslo, 2001. [22] J.F. Hair, R.E. Anderson, R.L. Tatham, W.C. Black, Multivariate Data Analysis, 5th ed., Prentice Hall Inc., USA, 1998. [23] D. Kirsanov, O. Mednova, V. Vietoris, P.A. Kilmartin, A. Legin, Talanta 90 (2012) 109. [24] S. Wold, M. Sjöström, L. Eriksson, Chemometrics and Intelligent Laboratory Systems 58 (2001) 109. [25] S. Serneels, C. Croux, P. Filzmoser, P.J. Van Espen, Chemometrics and Intelligent Laboratory Systems 79 (2005) 55. [26] R Development Core Team, R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria, 2010http://www.R-project.orghttp://cran.r-project.org/web/packages/chemometrics/index.html. [27] http://cran.r-project.org/web/packages/chemometrics/index.html [28] M. Daszykowski, K. Kaczmarek, Y. Vander Heyden, B. Walczak, Chemometrics and Intelligent Laboratory Systems 85 (2007) 203. [29] R.G. Brereton, TrAC-Trends in Analytical Chemistry 25 (2006) 1103. [30] P. Filzmoser, B. Liebmann, K. Varmuza, Journal of Chemometrics 33 (2009) 160. [31] K.H. Esbensen, P. Geladi, Journal of Chemometrics 24 (2010) 168.
